ganpdfs 1.1.0.dev0

GANs for PDF replicas

GANPDFs

Enhance the statistics of a prior PDF set by generating fake PDF replicas using Generative Adversarial Neural Networks (GANs). Documentation is available at https://n3pdf.github.io/ganpdfs/.

How to install

To install the ganpdfs package, just type

python setup.py install or python setup.py develop (if you are a developper)

The package can be installed via the Python Package Index (PyPI) by running:

pip install ganpdfs --upgrade

How to run

The code requires as an input a runcard.yml file in which the name of the PDF set and the characteristics of the Neural Network Models are defined. Examples of runcards can be found in the runcard folder.

ganpdfs runcard/reference.yml [-t TOT_REPLICAS_SIZE]

In case one does not want to train the GANs and directly resort to a pre-trained one, a pre-trained model can be used out of the box by setting the entry use_saved_model to True in the runcard.

In order to evolve the generated output grids, just run:

evolven3fit <PRIOR_PDF_NAME>_enhanced <TOT_REPLICAS_SIZE>

Then, to link the generated PDF set to the LHAPDF data directory, use the postgans script by running:

postgans --pdf <PRIOR_PDF_NAME> --nenhanced <TOT_REPLICAS_SIZE>

Hyper-parameter opitmization

For more details on how to define specific parameters when running the code and on how to perform a hyper-parameter scan, please head to the section how to of the documentation.